Modular infrastructure separates search from generation
Perplexity is moving away from a single model interface to a tiered system of APIs following its $500 million funding round and $9 billion valuation. The company now separates its developer products into the Agent API, Orchestrator API, Search API, and Embeddings API. I find the separation of the Search API from the Sonar model tokens to be the most significant change for developers building retrieval-heavy applications. The Search API returns ranked web results without a written answer for $5 per 1,000 successful requests, or $1 per 1,000 for the Fast Search tier. This pricing structure allows developers to avoid token costs for retrieval-only tasks. One request containing up to five queries counts as one billing unit, and this is distinct from rate limiting where five queries consume five query units against a documented limit of 50 query units per second. I would avoid the standard Search API if you need a synthesized response, but it works well for building reranking or citation layers. The separation acknowledges that search, answer generation, and agents are not interchangeable workloads. Developers building research tools or shopping interfaces can use raw results to maintain control over hallucination-prone synthesis. The practical winner will be the vendor that makes retrieval quality, billing, and failure behavior predictable enough to embed in real products, rather than the vendor with the loudest model claims.
Sonar Pro costs and the Agent API complexity
The Sonar Pro launch forces a comparison between token-based generation and request-based retrieval. While the Search API is cheap for raw data, Sonar models charge per million tokens. For example, Sonar Pro costs $3.00 per million input tokens and $15.00 per million output tokens.
| Model | Input /1M | Output /1M | Context |
|---|---|---|---|
| Sonar | $1.00 | $1.00 | 127,072 |
| Sonar Pro | $3.00 | $15.00 | 200,000 |
| Sonar Reasoning Pro | $2.00 | $8.00 | 128,000 |
| Sonar Deep Research | $2.00 | $8.00 | 128,000 |
| Sonar Pro Search | $3.00 | $15.00 | 200,000 |
The Agent API combines these token costs with fees for tool use like web search or URL fetching. A single research run can become expensive if the model enters a loop or calls tools repeatedly. A representative "low" preset run using 2,000 input tokens and 1,000 output tokens with one Fast Search and one fetch costs $0.0055. I see the Agent API as the primary choice for workflows requiring grounding, while the Search API remains better for simple retrieval. Does the increased complexity of the Agent API billing justify the switch from the legacy Sonar Chat Completions? The company now labels Sonar Chat Completions as legacy and directs developers toward the Agent API. The official pricing documentation also lists a faster Search API tier at $1 per 1,000 requests. This tiered latency pricing suggests a trade-off between speed and cost rather than a single optimal endpoint.
Consumer plans and coding performance
Perplexity differentiates its consumer and enterprise tiers from programmatic API usage. A Pro subscription costs $20 per month or $17 per month when you bill annually. This plan includes unlimited Pro Search and 20 Deep Research runs per day. I recommend the Pro plan for most researchers because the $20 price point covers the search needs that the free tier limits to five Pro searches per day. For heavy users, the Max plan costs $200 per month and includes 10,000 Computer credits every month. Enterprise Pro costs $40 per seat per month and provides 500 Deep Research runs per day alongside SSO and SOC 2 compliance. Enterprise Max costs $325 per seat per month and includes 15,000 Computer credits per seat.
| Plan | Monthly Price | Annual Price | Key Feature |
|---|---|---|---|
| Free | $0 | $0 | 5 Pro searches/day |
| Pro | $20 | $17 | Unlimited Pro Search |
| Max | $200 | $167 | 10,000 Computer credits |
| Enterprise Pro | $40 | $40 | 500 Deep Research/day |
| Enterprise Max | $325 | $271 | 15,000 Computer credits |
The coding performance of Perplexity Pro is high, as it passed all four programming tests in my testing, specifically a complex WordPress plugin task that failed for Claude 3.5 Sonnet. The Pro version used GPT-4 to analyze the code and successfully identified a parameter calling error in a WordPress framework that other models missed. I noticed the interface worked well for the WordPress plugin and string function rewrites. However, the site lacks any way to set a password or multifactor authentication. This forces users to rely solely on email codes for sign-in. If you need to build a production-ready research pipeline, you should budget for separate API credits because enterprise seats do not include API consumption. You should also compare the total cost per completed task rather than just dollars per million tokens.
